End-to-end lightweight method and apparatus for license plate recognition

1. An end-to-end lightweight method for license plate recognition, comprising: obtaining an image to be recognized; obtaining a number of a license plate in the image to be recognized and position coordinates of the license plate in the image to be recognized on the basis of the image to be recognized and a pre-trained target license plate recognition model, wherein the target license plate recognition model comprises a target feature extraction network, a target region candidate localization network, a target super-resolution generation network and a target recurrent neural network, the process of obtaining a number of a license plate in the image to be recognized and position coordinates of the license plate in the image to be recognized comprising: inputting the image to be recognized into the target feature extraction network and performing a first preset convolution operation on the image to be recognized to obtain a feature map; inputting the feature map into the target region candidate localization network, and performing a second preset convolution operation on the feature map to obtain coordinates of a candidate box; inputting the coordinates of the candidate box and the feature map into the target super-resolution generation network, and generating a super-resolution license plate image based on the coordinates of the candidate box and the feature map; inputting the super-resolution license plate image and the coordinates of the candidate box into the target recurrent neural network, and obtaining the number of the license plate in the image to be recognized based on the super-resolution license plate image and correspondence between features of the license plate and characters contained in the target recurrent neural network, and outputting the number, the coordinates of the candidate box and the super-resolution license plate image.

2. The method of claim 1 , wherein inputting the coordinates of the candidate box and the feature map into the target super-resolution generation network and generating a super-resolution license plate image based on the coordinates of the candidate box and the feature map comprises: inputting the coordinates of the candidate box and the feature map into the target super-resolution generation network; extracting features of the license plate from a region in the feature map corresponding to the coordinates of the candidate box; generating a super-resolution license plate image based on the extracted features of the license plate.

3. The method of claim 1 , wherein the process of training the target license plate recognition model comprises: constructing an initial license plate recognition model, wherein the initial license plate recognition model comprises an initial feature extraction network, an initial region candidate localization network, an initial super-resolution generation network and an initial recurrent neural network; obtaining image samples, and recording position coordinates and number of the license plate in each of the image samples; inputting the image samples into the initial feature extraction network, and performing a first preset convolution operation on the image samples to obtain a feature map sample corresponding to each of the image samples; inputting the feature map sample corresponding to each of the image samples into the initial region candidate localization network to obtain predicted position coordinates; adjusting parameters of the initial feature extraction network and of the initial region candidate localization network based on differences between the recorded position coordinates and the predicted position coordinates, and obtaining the target feature extraction network and the target region candidate localization network when the number of iterations reaches a first preset number of times; inputting each of the image samples into the target feature extraction network to obtain a target feature map sample corresponding to each of the image samples, and inputting the target feature map samples into the target region candidate localization network to obtain target position coordinates; inputting the target feature map samples and the target position coordinates into the initial super-resolution generation network to obtain predicted super-resolution images; adjusting parameters of the initial super-resolution generation network based on the degree of differences between the predicted super-resolution images and real license plates, and obtaining the target super-resolution generation network when the number of iterations reaches a second preset number of times; inputting the target feature map sample and target position coordinates corresponding to each of the image samples into the target super-resolution generation network to obtain a super-resolution license plate image sample corresponding to each of the image samples; inputting the super-resolution license plate image samples into the initial recurrent neural network to obtain predicted license plate numbers; adjusting parameters of the initial recurrent neural network based on the degree of differences between the recorded numbers and the predicted license plate numbers, and finishing the training and obtaining the target recurrent neural network when the number of iterations reaches a third preset number of times.

4. The method of claim 2 , wherein the process of training the target license plate recognition model comprises: constructing an initial license plate recognition model, wherein the initial license plate recognition model comprises an initial feature extraction network, an initial region candidate localization network, an initial super-resolution generation network and an initial recurrent neural network; obtaining image samples, and recording position coordinates and number of the license plate in each of the image samples; inputting the image samples into the initial feature extraction network, and performing a first preset convolution operation on the image samples to obtain a feature map sample corresponding to each of the image samples; inputting the feature map sample corresponding to each of the image samples into the initial region candidate localization network to obtain predicted position coordinates; adjusting parameters of the initial feature extraction network and of the initial region candidate localization network based on differences between the recorded position coordinates and the predicted position coordinates, and obtaining the target feature extraction network and the target region candidate localization network when the number of iterations reaches a first preset number of times; inputting each of the image samples into the target feature extraction network to obtain a target feature map sample corresponding to each of the image samples, and inputting the target feature map samples into the target region candidate localization network to obtain target position coordinates; inputting the target feature map samples and the target position coordinates into the initial super-resolution generation network to obtain predicted super-resolution images; adjusting parameters of the initial super-resolution generation network based on the degree of differences between the predicted super-resolution images and real license plates, and obtaining the target super-resolution generation network when the number of iterations reaches a second preset number of times; inputting the target feature map sample and target position coordinates corresponding to each of the image samples into the target super-resolution generation network to obtain a super-resolution license plate image sample corresponding to each of the image samples; inputting the super-resolution license plate image samples into the initial recurrent neural network to obtain predicted license plate numbers; adjusting parameters of the initial recurrent neural network based on the degree of differences between the recorded numbers and the predicted license plate numbers, and finishing the training and obtaining the target recurrent neural network when the number of iterations reaches a third preset number of times.

5. The method of claim 3 , wherein before inputting the image samples into the initial feature extraction network, the method further comprises: adjusting sizes of the obtained image samples to obtain resized image samples; wherein inputting the image samples into the initial feature extraction network comprises: inputting the resized image samples into the initial feature extraction network.

6. The method of claim 1 , wherein the target feature extraction network is a lightweight feature extraction convolutional neural network.

7. The method of claim 2 , wherein the target feature extraction network is a lightweight feature extraction convolutional neural network.

8. The method of claim 3 , wherein the target feature extraction network is a lightweight feature extraction convolutional neural network.

9. The method of claim 4 , wherein the target feature extraction network is a lightweight feature extraction convolutional neural network.

10. The method of claim 5 , wherein the target feature extraction network is a lightweight feature extraction convolutional neural network.

11. An end-to-end lightweight apparatus for license plate recognition, comprising: an image to be recognized obtaining module, configured for obtaining an image to be recognized; a license plate recognition model, configured for obtaining a license plate number in the image to be recognized and position coordinates of the license plate in the image to be recognized on the basis of the image to be recognized and a target license plate recognition model that has been trained by a recognition model training model in advance, wherein the target license plate recognition model comprises a target feature extraction network, a target region candidate localization network, a target super-resolution generation network and a target recurrent neural network; wherein the license plate recognition model is specifically configured for inputting the image to be recognized into the target feature extraction network and performing a first preset convolution operation on the image to be recognized to obtain a feature map; inputting the feature map into the target region candidate localization network, and performing a second preset convolution operation on the feature map to obtain coordinates of a candidate box; inputting the coordinates of the candidate box and the feature map into the target super-resolution generation network, and generating a super-resolution license plate image based on the coordinates of the candidate box and the feature map; inputting the super-resolution license plate image and the coordinates of the candidate box into the target recurrent neural network, obtaining the number of the license plate in the image to be recognized based on the super-resolution license plate image and correspondence between features of the license plate and characters contained in the target recurrent neural network, and outputting the number, the coordinates of the candidate box and the super-resolution license plate image.

12. The apparatus of claim 11 , wherein the license plate recognition model comprises: a super-resolution image generating unit, configured for inputting the coordinates of the candidate box and the feature map into the target super-resolution generation network; extracting features of the license plate from a region in the feature map corresponding to the coordinates of the candidate box; generating a super-resolution license plate image based on the extracted features of the license plate.

13. The apparatus of claim 11 , wherein the recognition model training module comprises: a recognition model constructing unit, configured for constructing an initial license plate recognition model, wherein the initial license plate recognition model comprises an initial feature extraction network, an initial region candidate localization network, an initial super-resolution generation network and an initial recurrent neural network; an image sample obtaining unit, configured for obtaining image samples, and recording position coordinates and number of the license plate in each of the image samples; a feature map sample obtaining unit, configured for inputting the image samples into the initial feature extraction network, and performing a first preset convolution operation on the image samples to obtain a feature map sample corresponding to each of the image samples; a predicted position coordinate determining unit, configured for inputting the feature map sample corresponding to each of the image samples into the initial region candidate localization network to obtain predicted position coordinates; a first parameter adjusting unit, configured for adjusting parameters of the initial feature extraction network and of the initial region candidate localization network based on differences between the recorded position coordinates and the predicted position coordinates, and obtaining the target feature extraction network and the target region candidate localization network when the number of iterations reaches a first preset number of times; a super-resolution image obtaining unit, configured for inputting the target feature map sample and target position coordinates corresponding to each of the image samples into the target super-resolution generation network to obtain a super-resolution license plate image sample corresponding to each of the image samples; a second parameter adjusting unit, configured for adjusting parameters of the initial super-resolution generation network based on the degree of differences between the predicted super-resolution images and real license plates, and obtaining the target super-resolution generation network when the number of iterations reaches a second preset number of times; a super-resolution license plate image sample obtaining unit, configured for inputting the feature map sample corresponding to each of the image samples into the target super-resolution generation network to obtain a super-resolution license plate image sample corresponding to each of the image samples; a predicted license plate number determining unit, configured for inputting the super-resolution license plate image samples into the initial recurrent neural network to obtain predicted license plate numbers; a third parameter adjusting unit, configured for adjusting parameters of the initial recurrent neural network based on the degree of differences between the recorded numbers and the predicted license plate numbers, and finishing the training and obtaining the target recurrent neural network when the number of iterations reaches a third preset number of times.

14. The apparatus of claim 13 , wherein the apparatus further comprises: a size adjusting module, configured for adjusting sizes of the obtained image samples, before inputting the image samples into the initial feature extraction network, to obtain resized image samples; the feature map sample obtaining unit comprises: a sample inputting subunit, configured for inputting the resized image samples into the initial feature extraction network.

15. The apparatus of claim 11 , wherein the target feature extraction network is a lightweight feature extraction convolutional neural network.

16. The apparatus of claim 12 , wherein the target feature extraction network is a lightweight feature extraction convolutional neural network.

17. The apparatus of claim 13 , wherein the target feature extraction network is a lightweight feature extraction convolutional neural network.

18. The apparatus of claim 14 , wherein the target feature extraction network is a lightweight feature extraction convolutional neural network.